Polygon Object Annotation via Recurrent GNN Decoders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object instance segmentation methods are time-consuming due to the need for manual annotation, especially in complex imagery, and existing polygon representation models have limitations in annotation speed and output resolution.
Innovation Solution
The proposed method employs a CNN encoder and a recurrent decoder with a Graph Neural Network to predict and refine polygon representations of objects, utilizing reinforcement learning for improved annotation efficiency and output resolution, and incorporates an evaluator network for selecting optimal polygon annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual pixel-level annotation is performed, then annotation accuracy is improved, but annotation time increases significantly
Solution Approach 1:
The patent segments the annotation task from pixel-level operations to polygon-vertex operations. Instead of annotating each pixel individually, the system identifies and annotates only the critical vertices that define the polygon boundary, reducing the annotation burden while maintaining accuracy.
Solution Approach 2:
The system uses a recurrent neural network to generate multiple candidate polygon annotations, then employs an evaluator network to select the best candidate. This copying and evaluation approach automates the selection process, reducing manual annotation time while preserving accuracy through automated quality assessment.
2Productivity
If polygon representation is used instead of pixel-level segmentation, then annotation speed is improved, but output resolution may be reduced
Solution Approach 1:
The system dynamically adjusts the number and position of polygon vertices based on the object complexity. The recurrent neural network can generate variable-length polygon sequences, adapting the resolution to the specific object being annotated rather than using a fixed vertex count, thus maintaining both speed and precision.
Solution Approach 2:
The patent replaces traditional mechanical polygon drawing interfaces with an automated recurrent neural network system that predicts polygon vertices. This substitution maintains high annotation speed while improving precision through learned patterns from training data, eliminating the trade-off between manual speed and automated accuracy.
3Measurement precision
If reinforcement learning is used for fine-tuning, then annotation accuracy is improved, but training complexity increases
Solution Approach 1:
The system performs preliminary supervised fine-tuning before reinforcement learning. This two-stage approach first establishes a solid baseline using supervised learning with labeled data, then applies reinforcement learning for further optimization. This preliminary action reduces the complexity burden by preparing the model in advance, making the subsequent reinforcement learning phase more efficient and manageable.
Solution Approach 2:
The evaluator network provides continuous feedback during the reinforcement learning process, assessing the quality of generated polygon annotations and guiding the policy gradient updates. This feedback mechanism structures the complex training process into manageable iterations, improving accuracy while organizing the training complexity into a systematic framework.
Data Source
AI summary
The present invention relates generally to object annotation, specifically to polygonal annotations of objects. Described are methods of annotating an object including steps of receiving an image depicting an object, generating a set of image features using a CNN encoder implemented on one or more computers, and producing a polygon object annotation via a recurrent decoder or a Graph Neural Network. The recurrent decoder may include a recurrent neural network, a graph neural network or a gated graph neural network. A system for annotating an object and a method of training an object annotation system are also described.